Anthropic, OpenAI, and Google APIs return encrypted chain-of-thought blocks. Researchers have found a method to decode this hidden reasoning at scale.
This matters because chain-of-thought is supposed to be opaque. If the encryption is exploitable, every intermediate reasoning step your model takes is potentially readable by third parties. That is not a theoretical risk.
The original piece details the specific exploit mechanism, which APIs are most vulnerable, and what scale means in practice. Read it.
[READ ORIGINAL →]